Object Detection
By Prof. Seungchul Lee
http://iai.postech.ac.kr/
Industrial AI Lab at POSTECH
http://iai.postech.ac.kr/
Industrial AI Lab at POSTECH
Table of Contents
Contents are retireved from
1. 2D Convolution¶
tf.keras.layers.Conv2D(filters, kernel_size, strides, padding, activation, kernel_regularizer, input_shape)
filters = 32
kernel_size = (3,3)
strides = (1,1)
padding = 'SAME'
activeation='relu'
kernel_regularizer=tf.keras.regularizers.l2(0.04)
input_shape = tensor of shape([input_h, input_w, input_ch])
- filter size
- the number of channels.
kernel_size
- the height and width of the 2D convolution window.
stride
- the step size of the kernel when traversing the image.
padding
- how the border of a sample is handled.
- A padded convolution will keep the spatial output dimensions equal to the input, whereas unpadded convolutions will crop away some of the borders if the kernel is larger than 1.
'SAME': enable zero padding'VALID': disable zero padding
- activation
- Activation function to use.
kernel_regularizer
- Initializer for the kernel weights matrix.
input and output channels
- A convolutional layer takes a certain number of input channels ($C$) and calculates a specific number of output channels ($D$).
Examples
input = [None, 4, 4, 1]
filter size = [3, 3, 1, 1]
strides = [1, 1, 1, 1]
padding = 'VALID'
input = [None, 5, 5, 1]
filter size = [3, 3, 1, 1]
strides = [1, 1, 1, 1]
padding = 'SAME'
2. Object Detection¶
2.1. Localization Methods¶
- Histogram of oriented Gradients (HOG) with SVM
- Selective search
3.1. One-stage Object Detection¶
- YOLO
- SSD
3.2. Two-stage Object Detection¶
- R-CNN
- Faster R-CNN
- Mask R-CNN
4. Examples¶
In [1]:
%%html
<center><iframe
width="560" height="315" src="https://www.youtube.com/embed/Cgxsv1riJhI" frameborder="0" allowfullscreen>
</iframe><center>
In [2]:
%%html
<center><iframe
width="560" height="315" src="https://www.youtube.com/embed/vRqSO6RsptU" frameborder="0">
</iframe><center>
5. Object Detection with Machinery Parts Dataset¶
- Simplified version of two-stage object detection model for tutorial
- 2-D convolution layers extract features from the input image.
- Extracted features are utilized for object bounding box detection and object classification
- Both classifier and bounding box regressor share the same features acuired from the 2-D convolution layers
- Dataset is available at: http://wordpress.csc.liv.ac.uk/smartlab/objectdetectionwork/
5.1. Import Library¶
In [1]:
import tensorflow as tf
import numpy as np
import matplotlib.pyplot as plt
import matplotlib.patches as patches
%matplotlib inline
5.2. Load Machinery Parts Dataset¶
Download data from here.
In [2]:
train_imgs = np.load('data_files/object_detction_trn_data.npy')
train_labels = np.load('data_files/object_detction_trn_label.npy')
test_imgs = np.load('data_files/object_detction_eval_data.npy')
test_labels = np.load('data_files/object_detction_eval_label.npy')
# input image: 240 by 320
# output label: class, x, y, h, w
classes = ['Axis',
'Bearing',
'Bearing_Box',
'Distance_Tube',
'F20_20_B']
In [3]:
print(train_imgs.shape)
print(train_labels.shape)
print(test_imgs.shape)
print(test_labels.shape)
Five classes images are availabe: Axis, bearing, bearing box, distance tube, beam
250 images are used for training (50 images per class)
50 images are avalialbe for evaluation (10 images per class)
One object per image (240 by 320)
Labeled with class and bounding box location(normalizsed): class, $x, y, h, \omega$
In [4]:
idx = 138
train_img = train_imgs[idx]
c, x, y, h, w = train_labels[idx]
# rescaling
x, w = x*320, w*320
y, h = y*240, h*240
rect = patches.Rectangle((x, y),
w,
h,
linewidth = 2,
edgecolor = 'r',
facecolor = 'none')
fig, ax = plt.subplots(figsize = (8,8))
plt.title(classes[int(c)])
plt.imshow(train_img)
ax.add_patch(rect)
plt.axis('off')
plt.show()
In [5]:
# rescaling output labels
train_labels = np.multiply(train_labels, [1, 320, 240, 320, 240])
test_labels = np.multiply(test_labels, [1, 320, 240, 320, 240])
In [6]:
feature_extractor = tf.keras.models.Sequential([
tf.keras.layers.Conv2D(filters = 32,
kernel_size = (3,3),
activation = 'relu',
padding = 'SAME',
input_shape = (240, 320, 3)),
tf.keras.layers.MaxPool2D(pool_size = (2,2)),
tf.keras.layers.Conv2D(64, (3,3), activation = 'relu', padding = 'SAME'),
tf.keras.layers.MaxPool2D((2,2)),
tf.keras.layers.Conv2D(64, (3,3), activation = 'relu', padding = 'SAME'),
tf.keras.layers.MaxPool2D((2,2)),
tf.keras.layers.Conv2D(128, (3,3), activation = 'relu', padding = 'SAME'),
tf.keras.layers.MaxPool2D((2,2)),
tf.keras.layers.Conv2D(128, (3,3), activation = 'relu', padding = 'SAME'),
tf.keras.layers.MaxPool2D((2,2)),
tf.keras.layers.Conv2D(256, (3,3), activation = 'relu', padding = 'SAME'),
tf.keras.layers.GlobalAveragePooling2D()
])
In [7]:
classifier = tf.keras.layers.Dense(256, activation = 'relu')(feature_extractor.output)
classifier = tf.keras.layers.Dense(256, activation = 'relu')(classifier)
classifier = tf.keras.layers.Dense(5, activation = 'softmax', name = 'cls')(classifier)
In [8]:
bb_regressor = tf.keras.layers.Dense(256, activation = 'relu')(feature_extractor.output)
bb_regressor = tf.keras.layers.Dense(256, activation = 'relu')(bb_regressor)
bb_regressor = tf.keras.layers.Dense(4, name = 'bbox')(bb_regressor)
In [9]:
object_detection = tf.keras.models.Model(inputs = feature_extractor.input,
outputs = [classifier, bb_regressor])
In [10]:
object_detection.summary()
5.4. Define Losses and Optimization Configuration¶
In [11]:
object_detection.compile(optimizer = 'adam',
loss = {'cls': 'sparse_categorical_crossentropy',
'bbox': 'mse'},
loss_weights = {'cls': 1,
'bbox': 1})
In [12]:
# divide labels to cls and bbox labels
train_cls = train_labels[:,:1]
train_bbox = train_labels[:,1:]
print(train_labels.shape)
print(train_cls.shape)
print(train_bbox.shape)
In [13]:
object_detection.fit(x = train_imgs,
y = {'cls': train_cls, 'bbox': train_bbox},
epochs = 40)
Out[13]:
5.5. Results for Training Sets¶
In [18]:
idx = 110
# true label
c_label, x_label, y_label, h_label, w_label = train_labels[idx]
rect_label = patches.Rectangle((x_label, y_label),
w_label,
h_label,
linewidth = 2,
edgecolor = 'r',
facecolor = 'none')
# predict
c_pred, bbox = object_detection(train_imgs[[idx]])
x, y, h, w = bbox[0]
rect = patches.Rectangle((x, y),
w,
h,
linewidth = 2,
edgecolor = 'b',
facecolor = 'none')
In [20]:
print(classes[int(c_label)])
print(classes[np.argmax(c_pred)])
In [21]:
fig, ax = plt.subplots(figsize = (8,8))
plt.imshow(train_imgs[idx])
ax.add_patch(rect_label)
ax.add_patch(rect)
plt.axis('off')
plt.show()
5.6. Results for Testing Sets¶
In [23]:
idx = 40
# true label
c_label, x_label, y_label, h_label, w_label = test_labels[idx]
rect_label = patches.Rectangle((x_label, y_label),
w_label,
h_label,
linewidth = 2,
edgecolor = 'r',
facecolor = 'none')
# predict
c_pred, bbox = object_detection(test_imgs[[idx]])
x, y, h, w = bbox[0]
rect = patches.Rectangle((x, y),
w,
h,
linewidth = 2,
edgecolor = 'b',
facecolor = 'none')
In [24]:
print(classes[int(c_label)])
print(classes[np.argmax(c_pred)])
In [25]:
fig, ax = plt.subplots(figsize = (8,8))
plt.imshow(test_imgs[idx])
ax.add_patch(rect_label)
ax.add_patch(rect)
plt.axis('off')
plt.show()
In [ ]:
%%javascript
$.getScript('https://kmahelona.github.io/ipython_notebook_goodies/ipython_notebook_toc.js')